A Nonlinear Dimensionality Reduction Method for Localized Tangent Space Alignment of High-dimensional Spectra of Tobacco Leaves

Regarding the problems of distance failure and dimensional disaster of tobacco high-dimensional near-infrared spectral similarity metrics during the cigarette recipe maintenance process, we propose an improved local tangent space alignment algorithm based on adaptive neighborhood preservation and stochastic projection optimization (PAG-LTSA). It significantly improves the computational efficiency of the algorithm through an adaptive neighborhood selection strategy based on distance probability equations and geodesic distances, as well as parallel accelerated computation and stochastic projection techniques. We verify the effectiveness of the PAG-LTSA algorithm on the actual production data of a beer enterprise. The results show that the dimensionality reduction results of PAG-LTSA algorithm have the most obvious discrimination. In the raw material origin classification experiment, the average accuracy of the algorithm is 9.58% higher than the traditional method, while the execution time is only 83.3% of the original one, which proves the effectiveness of the algorithm.